Skip to content
LIAW_2024_05 // THERM:03

Thermal Loads in High-Density Stacks

Packing normalization cores tight enough to win the latency race without cooking the freshness guarantees.

Co-location wins latency, and density wins co-location. The closer the normalization cores sit to the ingest NICs, the fewer microseconds a tick spends in transit — so we pack the stack tight. The price of that proximity is heat, and heat is a freshness problem wearing a hardware costume.

When a dense stack runs hot, the firmware throttles. Throttling does not crash anything; it just quietly stretches every per-event deadline. In a naive system that is invisible until a desk complains. In LIAW it surfaces immediately, because the provenance stamp records observed processing time, and a thermal slowdown shows up as a measurable drift in per-source freshness.

So the cooling policy is written against the latency budget, not against a temperature alarm. We cap sustained clock before we approach the thermal ceiling, trading a sliver of peak throughput for a flat freshness curve. A predictable slow core beats an occasionally-fast one when consumers are paying for determinism.

TEXT
STACK     CORES   SUSTAINED_GHZ   THERMAL_HEADROOM   FRESHNESS_P99
rack-a1   64      3.4             18 C               0.05 ms
rack-a2   64      3.4             21 C               0.05 ms
rack-b1   48      3.6             14 C               0.06 ms

The result is boring on purpose: a freshness curve you can plot as a flat line, because the thermal envelope was budgeted before the first core was ever racked.

§ · CONTINUE THE EXPLORATION

The ledger is always open.

Read the source-aware provenance trail, or pull the full dispatch for offline review.